app.py

import tqdm
from PIL import Image
import hashlib
import torch
import torch.nn.functional as F
import fitz
import threading
import gradio as gr
import spaces
import os
from transformers import AutoModel
from transformers import AutoTokenizer
from PIL import Image
import torch
import os
import numpy as np
import json
import time

cache_dir = '/data/KB'
os.makedirs(cache_dir, exist_ok=True)

@spaces.GPU
def weighted_mean_pooling(hidden, attention_mask):
    attention_mask_ = attention_mask * attention_mask.cumsum(dim=1)
    s = torch.sum(hidden * attention_mask_.unsqueeze(-1).float(), dim=1)
    d = attention_mask_.sum(dim=1, keepdim=True).float()
    reps = s / d
    return reps

@spaces.GPU
@torch.no_grad()
def encode(text_or_image_list):
    global model, tokenizer
    if (isinstance(text_or_image_list[0], str)):
        inputs = {
            "text": text_or_image_list,
            'image': [None] * len(text_or_image_list),
            'tokenizer': tokenizer
        }
    else:
        inputs = {
            "text": [''] * len(text_or_image_list),
            'image': text_or_image_list,
            'tokenizer': tokenizer
        }
    outputs = model(**inputs)
    attention_mask = outputs.attention_mask
    hidden = outputs.last_hidden_state

    reps = weighted_mean_pooling(hidden, attention_mask)   
    embeddings = F.normalize(reps, p=2, dim=1).detach().cpu().numpy()
    return embeddings

@spaces.GPU
def add_pdf_gradio(pdf_file_list, progress=gr.Progress()):
    global model, tokenizer
    model.eval()

    print(pdf_file_list)
    
    pdf_file_list = sorted(pdf_file_list, key=lambda x: os.path.basename(x))
    print(pdf_file_list)
    knowledge_base_name = str(int(time.time()))
    this_cache_dir = os.path.join(cache_dir, knowledge_base_name)
    os.makedirs(this_cache_dir, exist_ok=True)
    index2img_filename = []

    for pdf_file_path in pdf_file_list:
        with open(os.path.join(this_cache_dir, os.path.basename(pdf_file_path)), 'wb') as file1:
            with open(pdf_file_path, "rb") as file2:
                file1.write(file2.read())

    for pdf_file_path in pdf_file_list:

        print(f"Processing {pdf_file_path}")

        pdf_name = os.path.basename(pdf_file_path)

        dpi = 200
        doc = fitz.open(pdf_file_path)
        
        reps_list = []
        images = []

        for page in progress.tqdm(doc):
            # with self.lock: # because we hope one 16G gpu only process one image at the same time
            pix = page.get_pixmap(dpi=dpi)
            image = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
            with torch.no_grad():
                reps = encode([image])
            reps_list.append(reps)
            images.append(image)

        for idx in range(len(images)):
            image = images[idx]
            cache_image_path = os.path.join(this_cache_dir, f"{pdf_name}_{idx}.png")
            image.save(cache_image_path)
            index2img_filename.append(os.path.basename(cache_image_path))

        np.save(os.path.join(this_cache_dir, f"{pdf_name.split('.')[0]}.npy"), reps_list)

    with open(os.path.join(this_cache_dir, f"index2img_filename.txt"), 'w') as f:
        f.write('\n'.join(index2img_filename))
        
    return knowledge_base_name

@spaces.GPU
def retrieve_gradio(knowledge_base: str, query: str, topk: int):
    global model, tokenizer

    model.eval()

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    if not os.path.exists(target_cache_dir):
        return None
    
    with open(os.path.join(target_cache_dir, f"index2img_filename.txt"), 'r') as f:
        index2img_filename = f.read().split('\n')
    
    doc_list = [f for f in os.listdir(target_cache_dir) if f.endswith('.npy')]
    doc_list = sorted(doc_list)
    doc_reps = [np.load(os.path.join(target_cache_dir, f)) for f in doc_list]
    doc_reps_cat = torch.cat([torch.Tensor(i) for i in doc_reps], dim=0)
    doc_reps_cat = torch.cat([i for i in doc_reps_cat], dim=0)

    query_with_instruction = "Represent this query for retrieving relevant document: " + query
    with torch.no_grad():
        query_rep = torch.Tensor(encode([query_with_instruction]))

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    print(f"query_rep_shape: {query_rep.shape}, doc_reps_cat_shape: {doc_reps_cat.shape}")
    similarities = torch.matmul(query_rep, doc_reps_cat.T)

    topk_values, topk_doc_ids = torch.topk(similarities, k=topk)

    topk_values_np = topk_values.squeeze(0).cpu().numpy()

    topk_doc_ids_np = topk_doc_ids.squeeze(0).cpu().numpy()

    similarities_np = similarities.cpu().numpy()
    print(f"topk_doc_ids_np: {topk_doc_ids_np}, topk_values_np: {topk_values_np}")
    images_topk = [Image.open(os.path.join(target_cache_dir, f"{index2img_filename[idx]}")) for idx in topk_doc_ids_np]

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'w') as f:
        f.write(json.dumps(
            {
                "knowledge_base": knowledge_base,
                "query": query,
                "retrived_docs": [os.path.join(target_cache_dir, f"{index2img_filename[idx]}") for idx in topk_doc_ids_np]
            }, indent=4, ensure_ascii=False
        ))

    return images_topk


def upvote(knowledge_base, query):
    global model, tokenizer

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
        data = json.loads(f.read())

    data["user_preference"] = "upvote"

    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
        f.write(json.dumps(data, indent=4, ensure_ascii=False))

    print("up", os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"))

    gr.Info('Received! Thank you very much!')

    return


def downvote(knowledge_base, query):
    global model, tokenizer

    target_cache_dir = os.path.join(cache_dir, knowledge_base)

    query_md5 = hashlib.md5(query.encode()).hexdigest()

    with open(os.path.join(target_cache_dir, f"q-{query_md5}.json"), 'r') as f:
        data = json.loads(f.read())

    data["user_preference"] = "downvote"

    with open(os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"), 'w') as f:
        f.write(json.dumps(data, indent=4, ensure_ascii=False))

    print("down", os.path.join(target_cache_dir, f"q-{query_md5}-withpref.json"))

    gr.Info('Received! Thank you very much!')

    return



device = 'cuda'

print("emb model load begin...")
model_path = 'openbmb/VisRAG-Ret' # replace with your local model path
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
model.eval()
model.to(device)
print("emb model load success!")

print("gen model load begin...")
gen_model_path = 'openbmb/MiniCPM-V-2_6'
gen_tokenizer = AutoTokenizer.from_pretrained(gen_model_path, attn_implementation='sdpa', trust_remote_code=True)
gen_model = AutoModel.from_pretrained(gen_model_path, trust_remote_code=True,
    attn_implementation='sdpa', torch_dtype=torch.bfloat16)
gen_model.eval()
gen_model.to(device)
print("gen model load success!")


@spaces.GPU
def answer_question(images, question):
    global gen_model, gen_tokenizer
    # here each element of images is a tuple of (image_path, None).
    images_ = [Image.open(image[0]).convert('RGB') for image in images]
    msgs = [{'role': 'user', 'content': [question, *images_]}]
    answer = gen_model.chat(
        image=None,
        msgs=msgs,
        tokenizer=gen_tokenizer
    )
    print(answer)
    return answer


with gr.Blocks() as app:
    gr.Markdown("# VisRAG Pipeline: Vision-based Retrieval-augmented Generation on Multi-modality Documents")
    
    gr.Markdown("""
- A Vision Language Model Dense Retriever ([VisRAG-Ret](https://huggingface.co/openbmb/VisRAG-Ret)) **directly reads** your PDFs **without need for OCR**, generates **multimodal dense representations** and assists in building your personal library. 

- **Ask a question**, and it will retrieve the most relevant pages. Then, [MiniCPM-V-2.6](https://huggingface.co/spaces/openbmb/MiniCPM-V-2_6) will answer your question based on the recalled pages, utilizing its strong multi-image understanding capabilities.

    - It assists you in reading **lengthy**, **visually-intensive** or **text-oriented** PDF documents, helping you locate pages that answer your questions.

    - It enables you to build a personal library and retrieve book pages from a large collection of books.

    - It works like a human: reading, storing, retrieving, and answering with full visual comprehension.
""")
    
    gr.Markdown("- The current online demo supports PDF documents with fewer than 50 pages due to GPU time limitations. For longer PDFs and books, consider deploying it on your own machine.")
    gr.Markdown("Thank you very much to [@bokesyo](https://huggingface.co/bokesyo) for writing the code.")
    
    with gr.Row():
        file_input = gr.File(file_types=["pdf"], file_count="multiple", label="Step 1: Upload PDF")
        file_result = gr.Text(label="Knowledge Base ID (remember it, it is re-usable!)")
        process_button = gr.Button("Process PDF (Don't click until PDF uploaded successfully)")
    
    process_button.click(add_pdf_gradio, inputs=[file_input], outputs=file_result)

    with gr.Row():
        kb_id_input = gr.Text(label="Your Knowledge Base ID (paste your Knowledge Base ID here, it is re-usable):")
        query_input = gr.Text(label="Your Queston")
        topk_input = inputs=gr.Number(value=1, minimum=1, maximum=10, step=1, label="Number of pages to retrieve")
        retrieve_button = gr.Button("Step2: Retrieve Pages")

    with gr.Row():
        gr.Examples(
            examples=[
                [["car_owner_manual.pdf"], "1731341207", "怀孕如何系安全带?"],
                [["car_owner_manual.pdf"], "1731341207", "什么时候会触发侧气囊弹出?"],
                [["main_figure.pdf"], "1731341098", "What is RAG-V?"],
                [["main_figure.pdf"], "1731341098", "How does RAG-V perform?"]
            ],
            inputs=[file_input, kb_id_input, query_input],
        )

    with gr.Row():
        images_output = gr.Gallery(label="Retrieved Pages")
    
    retrieve_button.click(retrieve_gradio, inputs=[kb_id_input, query_input, topk_input], outputs=images_output)

    with gr.Row():
        button = gr.Button("Step 3: Answer Question with Retrieved Pages")

        gen_model_response = gr.Textbox(label="MiniCPM-V-2.6's Answer")

        button.click(fn=answer_question, inputs=[images_output, query_input], outputs=gen_model_response)
    
    with gr.Row():
        downvote_button = gr.Button("🤣Downvote")
        upvote_button = gr.Button("🤗Upvote")
        
    upvote_button.click(upvote, inputs=[kb_id_input, query_input], outputs=None)
    downvote_button.click(downvote, inputs=[kb_id_input, query_input], outputs=None)

    gr.Markdown("By using this demo, you agree to share your usage data with us for research purposes, helping us improve the user experience.")


app.launch()